1 /*- 2 * SPDX-License-Identifier: BSD-4-Clause 3 * 4 * Copyright (c) 1998 Matthew Dillon, 5 * Copyright (c) 1994 John S. Dyson 6 * Copyright (c) 1990 University of Utah. 7 * Copyright (c) 1982, 1986, 1989, 1993 8 * The Regents of the University of California. All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * the Systems Programming Group of the University of Utah Computer 12 * Science Department. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * New Swap System 43 * Matthew Dillon 44 * 45 * Radix Bitmap 'blists'. 46 * 47 * - The new swapper uses the new radix bitmap code. This should scale 48 * to arbitrarily small or arbitrarily large swap spaces and an almost 49 * arbitrary degree of fragmentation. 50 * 51 * Features: 52 * 53 * - on the fly reallocation of swap during putpages. The new system 54 * does not try to keep previously allocated swap blocks for dirty 55 * pages. 56 * 57 * - on the fly deallocation of swap 58 * 59 * - No more garbage collection required. Unnecessarily allocated swap 60 * blocks only exist for dirty vm_page_t's now and these are already 61 * cycled (in a high-load system) by the pager. We also do on-the-fly 62 * removal of invalidated swap blocks when a page is destroyed 63 * or renamed. 64 * 65 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 66 * 67 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 68 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 69 */ 70 71 #include <sys/cdefs.h> 72 __FBSDID("$FreeBSD$"); 73 74 #include "opt_vm.h" 75 76 #include <sys/param.h> 77 #include <sys/bio.h> 78 #include <sys/blist.h> 79 #include <sys/buf.h> 80 #include <sys/conf.h> 81 #include <sys/disk.h> 82 #include <sys/disklabel.h> 83 #include <sys/eventhandler.h> 84 #include <sys/fcntl.h> 85 #include <sys/lock.h> 86 #include <sys/kernel.h> 87 #include <sys/mount.h> 88 #include <sys/namei.h> 89 #include <sys/malloc.h> 90 #include <sys/pctrie.h> 91 #include <sys/priv.h> 92 #include <sys/proc.h> 93 #include <sys/racct.h> 94 #include <sys/resource.h> 95 #include <sys/resourcevar.h> 96 #include <sys/rwlock.h> 97 #include <sys/sbuf.h> 98 #include <sys/sysctl.h> 99 #include <sys/sysproto.h> 100 #include <sys/systm.h> 101 #include <sys/sx.h> 102 #include <sys/vmmeter.h> 103 #include <sys/vnode.h> 104 105 #include <security/mac/mac_framework.h> 106 107 #include <vm/vm.h> 108 #include <vm/pmap.h> 109 #include <vm/vm_map.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_object.h> 112 #include <vm/vm_page.h> 113 #include <vm/vm_pager.h> 114 #include <vm/vm_pageout.h> 115 #include <vm/vm_param.h> 116 #include <vm/swap_pager.h> 117 #include <vm/vm_extern.h> 118 #include <vm/uma.h> 119 120 #include <geom/geom.h> 121 122 /* 123 * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. 124 * The 64-page limit is due to the radix code (kern/subr_blist.c). 125 */ 126 #ifndef MAX_PAGEOUT_CLUSTER 127 #define MAX_PAGEOUT_CLUSTER 32 128 #endif 129 130 #if !defined(SWB_NPAGES) 131 #define SWB_NPAGES MAX_PAGEOUT_CLUSTER 132 #endif 133 134 #define SWAP_META_PAGES PCTRIE_COUNT 135 136 /* 137 * A swblk structure maps each page index within a 138 * SWAP_META_PAGES-aligned and sized range to the address of an 139 * on-disk swap block (or SWAPBLK_NONE). The collection of these 140 * mappings for an entire vm object is implemented as a pc-trie. 141 */ 142 struct swblk { 143 vm_pindex_t p; 144 daddr_t d[SWAP_META_PAGES]; 145 }; 146 147 static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); 148 static struct mtx sw_dev_mtx; 149 static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); 150 static struct swdevt *swdevhd; /* Allocate from here next */ 151 static int nswapdev; /* Number of swap devices */ 152 int swap_pager_avail; 153 static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ 154 155 static __exclusive_cache_line u_long swap_reserved; 156 static u_long swap_total; 157 static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); 158 159 static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 160 "VM swap stats"); 161 162 SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 163 &swap_reserved, 0, sysctl_page_shift, "A", 164 "Amount of swap storage needed to back all allocated anonymous memory."); 165 SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, 166 &swap_total, 0, sysctl_page_shift, "A", 167 "Total amount of available swap storage."); 168 169 static int overcommit = 0; 170 SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, 171 "Configure virtual memory overcommit behavior. See tuning(7) " 172 "for details."); 173 static unsigned long swzone; 174 SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, 175 "Actual size of swap metadata zone"); 176 static unsigned long swap_maxpages; 177 SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, 178 "Maximum amount of swap supported"); 179 180 static COUNTER_U64_DEFINE_EARLY(swap_free_deferred); 181 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, 182 CTLFLAG_RD, &swap_free_deferred, 183 "Number of pages that deferred freeing swap space"); 184 185 static COUNTER_U64_DEFINE_EARLY(swap_free_completed); 186 SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, 187 CTLFLAG_RD, &swap_free_completed, 188 "Number of deferred frees completed"); 189 190 /* bits from overcommit */ 191 #define SWAP_RESERVE_FORCE_ON (1 << 0) 192 #define SWAP_RESERVE_RLIMIT_ON (1 << 1) 193 #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) 194 195 static int 196 sysctl_page_shift(SYSCTL_HANDLER_ARGS) 197 { 198 uint64_t newval; 199 u_long value = *(u_long *)arg1; 200 201 newval = ((uint64_t)value) << PAGE_SHIFT; 202 return (sysctl_handle_64(oidp, &newval, 0, req)); 203 } 204 205 static bool 206 swap_reserve_by_cred_rlimit(u_long pincr, struct ucred *cred, int oc) 207 { 208 struct uidinfo *uip; 209 u_long prev; 210 211 uip = cred->cr_ruidinfo; 212 213 prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); 214 if ((oc & SWAP_RESERVE_RLIMIT_ON) != 0 && 215 prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && 216 priv_check(curthread, PRIV_VM_SWAP_NORLIMIT) != 0) { 217 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); 218 KASSERT(prev >= pincr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 219 return (false); 220 } 221 return (true); 222 } 223 224 static void 225 swap_release_by_cred_rlimit(u_long pdecr, struct ucred *cred) 226 { 227 struct uidinfo *uip; 228 #ifdef INVARIANTS 229 u_long prev; 230 #endif 231 232 uip = cred->cr_ruidinfo; 233 234 #ifdef INVARIANTS 235 prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); 236 KASSERT(prev >= pdecr, ("negative vmsize for uid = %d\n", uip->ui_uid)); 237 #else 238 atomic_subtract_long(&uip->ui_vmsize, pdecr); 239 #endif 240 } 241 242 static void 243 swap_reserve_force_rlimit(u_long pincr, struct ucred *cred) 244 { 245 struct uidinfo *uip; 246 247 uip = cred->cr_ruidinfo; 248 atomic_add_long(&uip->ui_vmsize, pincr); 249 } 250 251 bool 252 swap_reserve(vm_ooffset_t incr) 253 { 254 255 return (swap_reserve_by_cred(incr, curthread->td_ucred)); 256 } 257 258 bool 259 swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) 260 { 261 u_long r, s, prev, pincr; 262 #ifdef RACCT 263 int error; 264 #endif 265 int oc; 266 static int curfail; 267 static struct timeval lastfail; 268 269 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 270 (uintmax_t)incr)); 271 272 #ifdef RACCT 273 if (RACCT_ENABLED()) { 274 PROC_LOCK(curproc); 275 error = racct_add(curproc, RACCT_SWAP, incr); 276 PROC_UNLOCK(curproc); 277 if (error != 0) 278 return (false); 279 } 280 #endif 281 282 pincr = atop(incr); 283 prev = atomic_fetchadd_long(&swap_reserved, pincr); 284 r = prev + pincr; 285 s = swap_total; 286 oc = atomic_load_int(&overcommit); 287 if (r > s && (oc & SWAP_RESERVE_ALLOW_NONWIRED) != 0) { 288 s += vm_cnt.v_page_count - vm_cnt.v_free_reserved - 289 vm_wire_count(); 290 } 291 if ((oc & SWAP_RESERVE_FORCE_ON) != 0 && r > s && 292 priv_check(curthread, PRIV_VM_SWAP_NOQUOTA) != 0) { 293 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 294 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 295 goto out_error; 296 } 297 298 if (!swap_reserve_by_cred_rlimit(pincr, cred, oc)) { 299 prev = atomic_fetchadd_long(&swap_reserved, -pincr); 300 KASSERT(prev >= pincr, ("swap_reserved < incr on overcommit fail")); 301 goto out_error; 302 } 303 304 return (true); 305 306 out_error: 307 if (ppsratecheck(&lastfail, &curfail, 1)) { 308 printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", 309 cred->cr_ruidinfo->ui_uid, curproc->p_pid, incr); 310 } 311 #ifdef RACCT 312 if (RACCT_ENABLED()) { 313 PROC_LOCK(curproc); 314 racct_sub(curproc, RACCT_SWAP, incr); 315 PROC_UNLOCK(curproc); 316 } 317 #endif 318 319 return (false); 320 } 321 322 void 323 swap_reserve_force(vm_ooffset_t incr) 324 { 325 u_long pincr; 326 327 KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, 328 (uintmax_t)incr)); 329 330 #ifdef RACCT 331 if (RACCT_ENABLED()) { 332 PROC_LOCK(curproc); 333 racct_add_force(curproc, RACCT_SWAP, incr); 334 PROC_UNLOCK(curproc); 335 } 336 #endif 337 pincr = atop(incr); 338 atomic_add_long(&swap_reserved, pincr); 339 swap_reserve_force_rlimit(pincr, curthread->td_ucred); 340 } 341 342 void 343 swap_release(vm_ooffset_t decr) 344 { 345 struct ucred *cred; 346 347 PROC_LOCK(curproc); 348 cred = curproc->p_ucred; 349 swap_release_by_cred(decr, cred); 350 PROC_UNLOCK(curproc); 351 } 352 353 void 354 swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) 355 { 356 u_long pdecr; 357 #ifdef INVARIANTS 358 u_long prev; 359 #endif 360 361 KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, 362 (uintmax_t)decr)); 363 364 pdecr = atop(decr); 365 #ifdef INVARIANTS 366 prev = atomic_fetchadd_long(&swap_reserved, -pdecr); 367 KASSERT(prev >= pdecr, ("swap_reserved < decr")); 368 #else 369 atomic_subtract_long(&swap_reserved, pdecr); 370 #endif 371 372 swap_release_by_cred_rlimit(pdecr, cred); 373 #ifdef RACCT 374 if (racct_enable) 375 racct_sub_cred(cred, RACCT_SWAP, decr); 376 #endif 377 } 378 379 static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ 380 static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ 381 static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ 382 static int nsw_wcount_async; /* limit async write buffers */ 383 static int nsw_wcount_async_max;/* assigned maximum */ 384 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 385 386 static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); 387 SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | 388 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", 389 "Maximum running async swap ops"); 390 static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); 391 SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | 392 CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", 393 "Swap Fragmentation Info"); 394 395 static struct sx sw_alloc_sx; 396 397 /* 398 * "named" and "unnamed" anon region objects. Try to reduce the overhead 399 * of searching a named list by hashing it just a little. 400 */ 401 402 #define NOBJLISTS 8 403 404 #define NOBJLIST(handle) \ 405 (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) 406 407 static struct pagerlst swap_pager_object_list[NOBJLISTS]; 408 static uma_zone_t swwbuf_zone; 409 static uma_zone_t swrbuf_zone; 410 static uma_zone_t swblk_zone; 411 static uma_zone_t swpctrie_zone; 412 413 /* 414 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 415 * calls hooked from other parts of the VM system and do not appear here. 416 * (see vm/swap_pager.h). 417 */ 418 static vm_object_t 419 swap_pager_alloc(void *handle, vm_ooffset_t size, 420 vm_prot_t prot, vm_ooffset_t offset, struct ucred *); 421 static void swap_pager_dealloc(vm_object_t object); 422 static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, 423 int *); 424 static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, 425 int *, pgo_getpages_iodone_t, void *); 426 static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); 427 static boolean_t 428 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); 429 static void swap_pager_init(void); 430 static void swap_pager_unswapped(vm_page_t); 431 static void swap_pager_swapoff(struct swdevt *sp); 432 static void swap_pager_update_writecount(vm_object_t object, 433 vm_offset_t start, vm_offset_t end); 434 static void swap_pager_release_writecount(vm_object_t object, 435 vm_offset_t start, vm_offset_t end); 436 437 struct pagerops swappagerops = { 438 .pgo_init = swap_pager_init, /* early system initialization of pager */ 439 .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ 440 .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 441 .pgo_getpages = swap_pager_getpages, /* pagein */ 442 .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ 443 .pgo_putpages = swap_pager_putpages, /* pageout */ 444 .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ 445 .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ 446 .pgo_update_writecount = swap_pager_update_writecount, 447 .pgo_release_writecount = swap_pager_release_writecount, 448 }; 449 450 /* 451 * swap_*() routines are externally accessible. swp_*() routines are 452 * internal. 453 */ 454 static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 455 static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 456 457 SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, 458 "Maximum size of a swap block in pages"); 459 460 static void swp_sizecheck(void); 461 static void swp_pager_async_iodone(struct buf *bp); 462 static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); 463 static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); 464 static int swapongeom(struct vnode *); 465 static int swaponvp(struct thread *, struct vnode *, u_long); 466 static int swapoff_one(struct swdevt *sp, struct ucred *cred); 467 468 /* 469 * Swap bitmap functions 470 */ 471 static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); 472 static daddr_t swp_pager_getswapspace(int *npages); 473 474 /* 475 * Metadata functions 476 */ 477 static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); 478 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 479 static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, 480 vm_pindex_t pindex, vm_pindex_t count); 481 static void swp_pager_meta_free_all(vm_object_t); 482 static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); 483 484 static void 485 swp_pager_init_freerange(daddr_t *start, daddr_t *num) 486 { 487 488 *start = SWAPBLK_NONE; 489 *num = 0; 490 } 491 492 static void 493 swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) 494 { 495 496 if (*start + *num == addr) { 497 (*num)++; 498 } else { 499 swp_pager_freeswapspace(*start, *num); 500 *start = addr; 501 *num = 1; 502 } 503 } 504 505 static void * 506 swblk_trie_alloc(struct pctrie *ptree) 507 { 508 509 return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? 510 M_USE_RESERVE : 0))); 511 } 512 513 static void 514 swblk_trie_free(struct pctrie *ptree, void *node) 515 { 516 517 uma_zfree(swpctrie_zone, node); 518 } 519 520 PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); 521 522 /* 523 * SWP_SIZECHECK() - update swap_pager_full indication 524 * 525 * update the swap_pager_almost_full indication and warn when we are 526 * about to run out of swap space, using lowat/hiwat hysteresis. 527 * 528 * Clear swap_pager_full ( task killing ) indication when lowat is met. 529 * 530 * No restrictions on call 531 * This routine may not block. 532 */ 533 static void 534 swp_sizecheck(void) 535 { 536 537 if (swap_pager_avail < nswap_lowat) { 538 if (swap_pager_almost_full == 0) { 539 printf("swap_pager: out of swap space\n"); 540 swap_pager_almost_full = 1; 541 } 542 } else { 543 swap_pager_full = 0; 544 if (swap_pager_avail > nswap_hiwat) 545 swap_pager_almost_full = 0; 546 } 547 } 548 549 /* 550 * SWAP_PAGER_INIT() - initialize the swap pager! 551 * 552 * Expected to be started from system init. NOTE: This code is run 553 * before much else so be careful what you depend on. Most of the VM 554 * system has yet to be initialized at this point. 555 */ 556 static void 557 swap_pager_init(void) 558 { 559 /* 560 * Initialize object lists 561 */ 562 int i; 563 564 for (i = 0; i < NOBJLISTS; ++i) 565 TAILQ_INIT(&swap_pager_object_list[i]); 566 mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); 567 sx_init(&sw_alloc_sx, "swspsx"); 568 sx_init(&swdev_syscall_lock, "swsysc"); 569 } 570 571 /* 572 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 573 * 574 * Expected to be started from pageout process once, prior to entering 575 * its main loop. 576 */ 577 void 578 swap_pager_swap_init(void) 579 { 580 unsigned long n, n2; 581 582 /* 583 * Number of in-transit swap bp operations. Don't 584 * exhaust the pbufs completely. Make sure we 585 * initialize workable values (0 will work for hysteresis 586 * but it isn't very efficient). 587 * 588 * The nsw_cluster_max is constrained by the bp->b_pages[] 589 * array, which has MAXPHYS / PAGE_SIZE entries, and our locally 590 * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 591 * constrained by the swap device interleave stripe size. 592 * 593 * Currently we hardwire nsw_wcount_async to 4. This limit is 594 * designed to prevent other I/O from having high latencies due to 595 * our pageout I/O. The value 4 works well for one or two active swap 596 * devices but is probably a little low if you have more. Even so, 597 * a higher value would probably generate only a limited improvement 598 * with three or four active swap devices since the system does not 599 * typically have to pageout at extreme bandwidths. We will want 600 * at least 2 per swap devices, and 4 is a pretty good value if you 601 * have one NFS swap device due to the command/ack latency over NFS. 602 * So it all works out pretty well. 603 */ 604 nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); 605 606 nsw_wcount_async = 4; 607 nsw_wcount_async_max = nsw_wcount_async; 608 mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); 609 610 swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); 611 swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); 612 613 /* 614 * Initialize our zone, taking the user's requested size or 615 * estimating the number we need based on the number of pages 616 * in the system. 617 */ 618 n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : 619 vm_cnt.v_page_count / 2; 620 swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, 621 pctrie_zone_init, NULL, UMA_ALIGN_PTR, 0); 622 if (swpctrie_zone == NULL) 623 panic("failed to create swap pctrie zone."); 624 swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, 625 NULL, NULL, _Alignof(struct swblk) - 1, 0); 626 if (swblk_zone == NULL) 627 panic("failed to create swap blk zone."); 628 n2 = n; 629 do { 630 if (uma_zone_reserve_kva(swblk_zone, n)) 631 break; 632 /* 633 * if the allocation failed, try a zone two thirds the 634 * size of the previous attempt. 635 */ 636 n -= ((n + 2) / 3); 637 } while (n > 0); 638 639 /* 640 * Often uma_zone_reserve_kva() cannot reserve exactly the 641 * requested size. Account for the difference when 642 * calculating swap_maxpages. 643 */ 644 n = uma_zone_get_max(swblk_zone); 645 646 if (n < n2) 647 printf("Swap blk zone entries changed from %lu to %lu.\n", 648 n2, n); 649 /* absolute maximum we can handle assuming 100% efficiency */ 650 swap_maxpages = n * SWAP_META_PAGES; 651 swzone = n * sizeof(struct swblk); 652 if (!uma_zone_reserve_kva(swpctrie_zone, n)) 653 printf("Cannot reserve swap pctrie zone, " 654 "reduce kern.maxswzone.\n"); 655 } 656 657 static vm_object_t 658 swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, 659 vm_ooffset_t offset) 660 { 661 vm_object_t object; 662 663 if (cred != NULL) { 664 if (!swap_reserve_by_cred(size, cred)) 665 return (NULL); 666 crhold(cred); 667 } 668 669 /* 670 * The un_pager.swp.swp_blks trie is initialized by 671 * vm_object_allocate() to ensure the correct order of 672 * visibility to other threads. 673 */ 674 object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + 675 PAGE_MASK + size)); 676 677 object->un_pager.swp.writemappings = 0; 678 object->handle = handle; 679 if (cred != NULL) { 680 object->cred = cred; 681 object->charge = size; 682 } 683 return (object); 684 } 685 686 /* 687 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 688 * its metadata structures. 689 * 690 * This routine is called from the mmap and fork code to create a new 691 * OBJT_SWAP object. 692 * 693 * This routine must ensure that no live duplicate is created for 694 * the named object request, which is protected against by 695 * holding the sw_alloc_sx lock in case handle != NULL. 696 */ 697 static vm_object_t 698 swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, 699 vm_ooffset_t offset, struct ucred *cred) 700 { 701 vm_object_t object; 702 703 if (handle != NULL) { 704 /* 705 * Reference existing named region or allocate new one. There 706 * should not be a race here against swp_pager_meta_build() 707 * as called from vm_page_remove() in regards to the lookup 708 * of the handle. 709 */ 710 sx_xlock(&sw_alloc_sx); 711 object = vm_pager_object_lookup(NOBJLIST(handle), handle); 712 if (object == NULL) { 713 object = swap_pager_alloc_init(handle, cred, size, 714 offset); 715 if (object != NULL) { 716 TAILQ_INSERT_TAIL(NOBJLIST(object->handle), 717 object, pager_object_list); 718 } 719 } 720 sx_xunlock(&sw_alloc_sx); 721 } else { 722 object = swap_pager_alloc_init(handle, cred, size, offset); 723 } 724 return (object); 725 } 726 727 /* 728 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 729 * 730 * The swap backing for the object is destroyed. The code is 731 * designed such that we can reinstantiate it later, but this 732 * routine is typically called only when the entire object is 733 * about to be destroyed. 734 * 735 * The object must be locked. 736 */ 737 static void 738 swap_pager_dealloc(vm_object_t object) 739 { 740 741 VM_OBJECT_ASSERT_WLOCKED(object); 742 KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); 743 744 /* 745 * Remove from list right away so lookups will fail if we block for 746 * pageout completion. 747 */ 748 if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { 749 VM_OBJECT_WUNLOCK(object); 750 sx_xlock(&sw_alloc_sx); 751 TAILQ_REMOVE(NOBJLIST(object->handle), object, 752 pager_object_list); 753 sx_xunlock(&sw_alloc_sx); 754 VM_OBJECT_WLOCK(object); 755 } 756 757 vm_object_pip_wait(object, "swpdea"); 758 759 /* 760 * Free all remaining metadata. We only bother to free it from 761 * the swap meta data. We do not attempt to free swapblk's still 762 * associated with vm_page_t's for this object. We do not care 763 * if paging is still in progress on some objects. 764 */ 765 swp_pager_meta_free_all(object); 766 object->handle = NULL; 767 object->type = OBJT_DEAD; 768 } 769 770 /************************************************************************ 771 * SWAP PAGER BITMAP ROUTINES * 772 ************************************************************************/ 773 774 /* 775 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 776 * 777 * Allocate swap for up to the requested number of pages. The 778 * starting swap block number (a page index) is returned or 779 * SWAPBLK_NONE if the allocation failed. 780 * 781 * Also has the side effect of advising that somebody made a mistake 782 * when they configured swap and didn't configure enough. 783 * 784 * This routine may not sleep. 785 * 786 * We allocate in round-robin fashion from the configured devices. 787 */ 788 static daddr_t 789 swp_pager_getswapspace(int *io_npages) 790 { 791 daddr_t blk; 792 struct swdevt *sp; 793 int mpages, npages; 794 795 KASSERT(*io_npages >= 1, 796 ("%s: npages not positive", __func__)); 797 blk = SWAPBLK_NONE; 798 mpages = *io_npages; 799 npages = imin(BLIST_MAX_ALLOC, mpages); 800 mtx_lock(&sw_dev_mtx); 801 sp = swdevhd; 802 while (!TAILQ_EMPTY(&swtailq)) { 803 if (sp == NULL) 804 sp = TAILQ_FIRST(&swtailq); 805 if ((sp->sw_flags & SW_CLOSING) == 0) 806 blk = blist_alloc(sp->sw_blist, &npages, mpages); 807 if (blk != SWAPBLK_NONE) 808 break; 809 sp = TAILQ_NEXT(sp, sw_list); 810 if (swdevhd == sp) { 811 if (npages == 1) 812 break; 813 mpages = npages - 1; 814 npages >>= 1; 815 } 816 } 817 if (blk != SWAPBLK_NONE) { 818 *io_npages = npages; 819 blk += sp->sw_first; 820 sp->sw_used += npages; 821 swap_pager_avail -= npages; 822 swp_sizecheck(); 823 swdevhd = TAILQ_NEXT(sp, sw_list); 824 } else { 825 if (swap_pager_full != 2) { 826 printf("swp_pager_getswapspace(%d): failed\n", 827 *io_npages); 828 swap_pager_full = 2; 829 swap_pager_almost_full = 1; 830 } 831 swdevhd = NULL; 832 } 833 mtx_unlock(&sw_dev_mtx); 834 return (blk); 835 } 836 837 static bool 838 swp_pager_isondev(daddr_t blk, struct swdevt *sp) 839 { 840 841 return (blk >= sp->sw_first && blk < sp->sw_end); 842 } 843 844 static void 845 swp_pager_strategy(struct buf *bp) 846 { 847 struct swdevt *sp; 848 849 mtx_lock(&sw_dev_mtx); 850 TAILQ_FOREACH(sp, &swtailq, sw_list) { 851 if (swp_pager_isondev(bp->b_blkno, sp)) { 852 mtx_unlock(&sw_dev_mtx); 853 if ((sp->sw_flags & SW_UNMAPPED) != 0 && 854 unmapped_buf_allowed) { 855 bp->b_data = unmapped_buf; 856 bp->b_offset = 0; 857 } else { 858 pmap_qenter((vm_offset_t)bp->b_data, 859 &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); 860 } 861 sp->sw_strategy(bp, sp); 862 return; 863 } 864 } 865 panic("Swapdev not found"); 866 } 867 868 869 /* 870 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 871 * 872 * This routine returns the specified swap blocks back to the bitmap. 873 * 874 * This routine may not sleep. 875 */ 876 static void 877 swp_pager_freeswapspace(daddr_t blk, daddr_t npages) 878 { 879 struct swdevt *sp; 880 881 if (npages == 0) 882 return; 883 mtx_lock(&sw_dev_mtx); 884 TAILQ_FOREACH(sp, &swtailq, sw_list) { 885 if (swp_pager_isondev(blk, sp)) { 886 sp->sw_used -= npages; 887 /* 888 * If we are attempting to stop swapping on 889 * this device, we don't want to mark any 890 * blocks free lest they be reused. 891 */ 892 if ((sp->sw_flags & SW_CLOSING) == 0) { 893 blist_free(sp->sw_blist, blk - sp->sw_first, 894 npages); 895 swap_pager_avail += npages; 896 swp_sizecheck(); 897 } 898 mtx_unlock(&sw_dev_mtx); 899 return; 900 } 901 } 902 panic("Swapdev not found"); 903 } 904 905 /* 906 * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats 907 */ 908 static int 909 sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) 910 { 911 struct sbuf sbuf; 912 struct swdevt *sp; 913 const char *devname; 914 int error; 915 916 error = sysctl_wire_old_buffer(req, 0); 917 if (error != 0) 918 return (error); 919 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 920 mtx_lock(&sw_dev_mtx); 921 TAILQ_FOREACH(sp, &swtailq, sw_list) { 922 if (vn_isdisk(sp->sw_vp, NULL)) 923 devname = devtoname(sp->sw_vp->v_rdev); 924 else 925 devname = "[file]"; 926 sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); 927 blist_stats(sp->sw_blist, &sbuf); 928 } 929 mtx_unlock(&sw_dev_mtx); 930 error = sbuf_finish(&sbuf); 931 sbuf_delete(&sbuf); 932 return (error); 933 } 934 935 /* 936 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 937 * range within an object. 938 * 939 * This is a globally accessible routine. 940 * 941 * This routine removes swapblk assignments from swap metadata. 942 * 943 * The external callers of this routine typically have already destroyed 944 * or renamed vm_page_t's associated with this range in the object so 945 * we should be ok. 946 * 947 * The object must be locked. 948 */ 949 void 950 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) 951 { 952 953 swp_pager_meta_free(object, start, size); 954 } 955 956 /* 957 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 958 * 959 * Assigns swap blocks to the specified range within the object. The 960 * swap blocks are not zeroed. Any previous swap assignment is destroyed. 961 * 962 * Returns 0 on success, -1 on failure. 963 */ 964 int 965 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 966 { 967 daddr_t addr, blk, n_free, s_free; 968 int i, j, n; 969 970 swp_pager_init_freerange(&s_free, &n_free); 971 VM_OBJECT_WLOCK(object); 972 for (i = 0; i < size; i += n) { 973 n = size - i; 974 blk = swp_pager_getswapspace(&n); 975 if (blk == SWAPBLK_NONE) { 976 swp_pager_meta_free(object, start, i); 977 VM_OBJECT_WUNLOCK(object); 978 return (-1); 979 } 980 for (j = 0; j < n; ++j) { 981 addr = swp_pager_meta_build(object, 982 start + i + j, blk + j); 983 if (addr != SWAPBLK_NONE) 984 swp_pager_update_freerange(&s_free, &n_free, 985 addr); 986 } 987 } 988 swp_pager_freeswapspace(s_free, n_free); 989 VM_OBJECT_WUNLOCK(object); 990 return (0); 991 } 992 993 static bool 994 swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, 995 vm_pindex_t pindex, daddr_t addr) 996 { 997 daddr_t dstaddr; 998 999 KASSERT(srcobject->type == OBJT_SWAP, 1000 ("%s: Srcobject not swappable", __func__)); 1001 if (dstobject->type == OBJT_SWAP && 1002 swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { 1003 /* Caller should destroy the source block. */ 1004 return (false); 1005 } 1006 1007 /* 1008 * Destination has no swapblk and is not resident, transfer source. 1009 * swp_pager_meta_build() can sleep. 1010 */ 1011 VM_OBJECT_WUNLOCK(srcobject); 1012 dstaddr = swp_pager_meta_build(dstobject, pindex, addr); 1013 KASSERT(dstaddr == SWAPBLK_NONE, 1014 ("Unexpected destination swapblk")); 1015 VM_OBJECT_WLOCK(srcobject); 1016 1017 return (true); 1018 } 1019 1020 /* 1021 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 1022 * and destroy the source. 1023 * 1024 * Copy any valid swapblks from the source to the destination. In 1025 * cases where both the source and destination have a valid swapblk, 1026 * we keep the destination's. 1027 * 1028 * This routine is allowed to sleep. It may sleep allocating metadata 1029 * indirectly through swp_pager_meta_build(). 1030 * 1031 * The source object contains no vm_page_t's (which is just as well) 1032 * 1033 * The source object is of type OBJT_SWAP. 1034 * 1035 * The source and destination objects must be locked. 1036 * Both object locks may temporarily be released. 1037 */ 1038 void 1039 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 1040 vm_pindex_t offset, int destroysource) 1041 { 1042 1043 VM_OBJECT_ASSERT_WLOCKED(srcobject); 1044 VM_OBJECT_ASSERT_WLOCKED(dstobject); 1045 1046 /* 1047 * If destroysource is set, we remove the source object from the 1048 * swap_pager internal queue now. 1049 */ 1050 if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && 1051 srcobject->handle != NULL) { 1052 VM_OBJECT_WUNLOCK(srcobject); 1053 VM_OBJECT_WUNLOCK(dstobject); 1054 sx_xlock(&sw_alloc_sx); 1055 TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, 1056 pager_object_list); 1057 sx_xunlock(&sw_alloc_sx); 1058 VM_OBJECT_WLOCK(dstobject); 1059 VM_OBJECT_WLOCK(srcobject); 1060 } 1061 1062 /* 1063 * Transfer source to destination. 1064 */ 1065 swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); 1066 1067 /* 1068 * Free left over swap blocks in source. 1069 * 1070 * We have to revert the type to OBJT_DEFAULT so we do not accidentally 1071 * double-remove the object from the swap queues. 1072 */ 1073 if (destroysource) { 1074 swp_pager_meta_free_all(srcobject); 1075 /* 1076 * Reverting the type is not necessary, the caller is going 1077 * to destroy srcobject directly, but I'm doing it here 1078 * for consistency since we've removed the object from its 1079 * queues. 1080 */ 1081 srcobject->type = OBJT_DEFAULT; 1082 } 1083 } 1084 1085 /* 1086 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 1087 * the requested page. 1088 * 1089 * We determine whether good backing store exists for the requested 1090 * page and return TRUE if it does, FALSE if it doesn't. 1091 * 1092 * If TRUE, we also try to determine how much valid, contiguous backing 1093 * store exists before and after the requested page. 1094 */ 1095 static boolean_t 1096 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, 1097 int *after) 1098 { 1099 daddr_t blk, blk0; 1100 int i; 1101 1102 VM_OBJECT_ASSERT_LOCKED(object); 1103 KASSERT(object->type == OBJT_SWAP, 1104 ("%s: object not swappable", __func__)); 1105 1106 /* 1107 * do we have good backing store at the requested index ? 1108 */ 1109 blk0 = swp_pager_meta_lookup(object, pindex); 1110 if (blk0 == SWAPBLK_NONE) { 1111 if (before) 1112 *before = 0; 1113 if (after) 1114 *after = 0; 1115 return (FALSE); 1116 } 1117 1118 /* 1119 * find backwards-looking contiguous good backing store 1120 */ 1121 if (before != NULL) { 1122 for (i = 1; i < SWB_NPAGES; i++) { 1123 if (i > pindex) 1124 break; 1125 blk = swp_pager_meta_lookup(object, pindex - i); 1126 if (blk != blk0 - i) 1127 break; 1128 } 1129 *before = i - 1; 1130 } 1131 1132 /* 1133 * find forward-looking contiguous good backing store 1134 */ 1135 if (after != NULL) { 1136 for (i = 1; i < SWB_NPAGES; i++) { 1137 blk = swp_pager_meta_lookup(object, pindex + i); 1138 if (blk != blk0 + i) 1139 break; 1140 } 1141 *after = i - 1; 1142 } 1143 return (TRUE); 1144 } 1145 1146 /* 1147 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 1148 * 1149 * This removes any associated swap backing store, whether valid or 1150 * not, from the page. 1151 * 1152 * This routine is typically called when a page is made dirty, at 1153 * which point any associated swap can be freed. MADV_FREE also 1154 * calls us in a special-case situation 1155 * 1156 * NOTE!!! If the page is clean and the swap was valid, the caller 1157 * should make the page dirty before calling this routine. This routine 1158 * does NOT change the m->dirty status of the page. Also: MADV_FREE 1159 * depends on it. 1160 * 1161 * This routine may not sleep. 1162 * 1163 * The object containing the page may be locked. 1164 */ 1165 static void 1166 swap_pager_unswapped(vm_page_t m) 1167 { 1168 struct swblk *sb; 1169 vm_object_t obj; 1170 1171 /* 1172 * Handle enqueing deferred frees first. If we do not have the 1173 * object lock we wait for the page daemon to clear the space. 1174 */ 1175 obj = m->object; 1176 if (!VM_OBJECT_WOWNED(obj)) { 1177 VM_PAGE_OBJECT_BUSY_ASSERT(m); 1178 /* 1179 * The caller is responsible for synchronization but we 1180 * will harmlessly handle races. This is typically provided 1181 * by only calling unswapped() when a page transitions from 1182 * clean to dirty. 1183 */ 1184 if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == 1185 PGA_SWAP_SPACE) { 1186 vm_page_aflag_set(m, PGA_SWAP_FREE); 1187 counter_u64_add(swap_free_deferred, 1); 1188 } 1189 return; 1190 } 1191 if ((m->a.flags & PGA_SWAP_FREE) != 0) 1192 counter_u64_add(swap_free_completed, 1); 1193 vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); 1194 1195 /* 1196 * The meta data only exists if the object is OBJT_SWAP 1197 * and even then might not be allocated yet. 1198 */ 1199 KASSERT(m->object->type == OBJT_SWAP, 1200 ("Free object not swappable")); 1201 1202 sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, 1203 rounddown(m->pindex, SWAP_META_PAGES)); 1204 if (sb == NULL) 1205 return; 1206 if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) 1207 return; 1208 swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); 1209 sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; 1210 swp_pager_free_empty_swblk(m->object, sb); 1211 } 1212 1213 /* 1214 * swap_pager_getpages() - bring pages in from swap 1215 * 1216 * Attempt to page in the pages in array "ma" of length "count". The 1217 * caller may optionally specify that additional pages preceding and 1218 * succeeding the specified range be paged in. The number of such pages 1219 * is returned in the "rbehind" and "rahead" parameters, and they will 1220 * be in the inactive queue upon return. 1221 * 1222 * The pages in "ma" must be busied and will remain busied upon return. 1223 */ 1224 static int 1225 swap_pager_getpages_locked(vm_object_t object, vm_page_t *ma, int count, 1226 int *rbehind, int *rahead) 1227 { 1228 struct buf *bp; 1229 vm_page_t bm, mpred, msucc, p; 1230 vm_pindex_t pindex; 1231 daddr_t blk; 1232 int i, maxahead, maxbehind, reqcount; 1233 1234 VM_OBJECT_ASSERT_WLOCKED(object); 1235 reqcount = count; 1236 1237 KASSERT(object->type == OBJT_SWAP, 1238 ("%s: object not swappable", __func__)); 1239 if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) { 1240 VM_OBJECT_WUNLOCK(object); 1241 return (VM_PAGER_FAIL); 1242 } 1243 1244 KASSERT(reqcount - 1 <= maxahead, 1245 ("page count %d extends beyond swap block", reqcount)); 1246 1247 /* 1248 * Do not transfer any pages other than those that are xbusied 1249 * when running during a split or collapse operation. This 1250 * prevents clustering from re-creating pages which are being 1251 * moved into another object. 1252 */ 1253 if ((object->flags & (OBJ_SPLIT | OBJ_DEAD)) != 0) { 1254 maxahead = reqcount - 1; 1255 maxbehind = 0; 1256 } 1257 1258 /* 1259 * Clip the readahead and readbehind ranges to exclude resident pages. 1260 */ 1261 if (rahead != NULL) { 1262 *rahead = imin(*rahead, maxahead - (reqcount - 1)); 1263 pindex = ma[reqcount - 1]->pindex; 1264 msucc = TAILQ_NEXT(ma[reqcount - 1], listq); 1265 if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) 1266 *rahead = msucc->pindex - pindex - 1; 1267 } 1268 if (rbehind != NULL) { 1269 *rbehind = imin(*rbehind, maxbehind); 1270 pindex = ma[0]->pindex; 1271 mpred = TAILQ_PREV(ma[0], pglist, listq); 1272 if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) 1273 *rbehind = pindex - mpred->pindex - 1; 1274 } 1275 1276 bm = ma[0]; 1277 for (i = 0; i < count; i++) 1278 ma[i]->oflags |= VPO_SWAPINPROG; 1279 1280 /* 1281 * Allocate readahead and readbehind pages. 1282 */ 1283 if (rbehind != NULL) { 1284 for (i = 1; i <= *rbehind; i++) { 1285 p = vm_page_alloc(object, ma[0]->pindex - i, 1286 VM_ALLOC_NORMAL); 1287 if (p == NULL) 1288 break; 1289 p->oflags |= VPO_SWAPINPROG; 1290 bm = p; 1291 } 1292 *rbehind = i - 1; 1293 } 1294 if (rahead != NULL) { 1295 for (i = 0; i < *rahead; i++) { 1296 p = vm_page_alloc(object, 1297 ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); 1298 if (p == NULL) 1299 break; 1300 p->oflags |= VPO_SWAPINPROG; 1301 } 1302 *rahead = i; 1303 } 1304 if (rbehind != NULL) 1305 count += *rbehind; 1306 if (rahead != NULL) 1307 count += *rahead; 1308 1309 vm_object_pip_add(object, count); 1310 1311 pindex = bm->pindex; 1312 blk = swp_pager_meta_lookup(object, pindex); 1313 KASSERT(blk != SWAPBLK_NONE, 1314 ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); 1315 1316 VM_OBJECT_WUNLOCK(object); 1317 bp = uma_zalloc(swrbuf_zone, M_WAITOK); 1318 /* Pages cannot leave the object while busy. */ 1319 for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { 1320 MPASS(p->pindex == bm->pindex + i); 1321 bp->b_pages[i] = p; 1322 } 1323 1324 bp->b_flags |= B_PAGING; 1325 bp->b_iocmd = BIO_READ; 1326 bp->b_iodone = swp_pager_async_iodone; 1327 bp->b_rcred = crhold(thread0.td_ucred); 1328 bp->b_wcred = crhold(thread0.td_ucred); 1329 bp->b_blkno = blk; 1330 bp->b_bcount = PAGE_SIZE * count; 1331 bp->b_bufsize = PAGE_SIZE * count; 1332 bp->b_npages = count; 1333 bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; 1334 bp->b_pgafter = rahead != NULL ? *rahead : 0; 1335 1336 VM_CNT_INC(v_swapin); 1337 VM_CNT_ADD(v_swappgsin, count); 1338 1339 /* 1340 * perform the I/O. NOTE!!! bp cannot be considered valid after 1341 * this point because we automatically release it on completion. 1342 * Instead, we look at the one page we are interested in which we 1343 * still hold a lock on even through the I/O completion. 1344 * 1345 * The other pages in our ma[] array are also released on completion, 1346 * so we cannot assume they are valid anymore either. 1347 * 1348 * NOTE: b_blkno is destroyed by the call to swapdev_strategy 1349 */ 1350 BUF_KERNPROC(bp); 1351 swp_pager_strategy(bp); 1352 1353 /* 1354 * Wait for the pages we want to complete. VPO_SWAPINPROG is always 1355 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1356 * is set in the metadata for each page in the request. 1357 */ 1358 VM_OBJECT_WLOCK(object); 1359 /* This could be implemented more efficiently with aflags */ 1360 while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { 1361 ma[0]->oflags |= VPO_SWAPSLEEP; 1362 VM_CNT_INC(v_intrans); 1363 if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, 1364 "swread", hz * 20)) { 1365 printf( 1366 "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", 1367 bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); 1368 } 1369 } 1370 VM_OBJECT_WUNLOCK(object); 1371 1372 /* 1373 * If we had an unrecoverable read error pages will not be valid. 1374 */ 1375 for (i = 0; i < reqcount; i++) 1376 if (ma[i]->valid != VM_PAGE_BITS_ALL) 1377 return (VM_PAGER_ERROR); 1378 1379 return (VM_PAGER_OK); 1380 1381 /* 1382 * A final note: in a low swap situation, we cannot deallocate swap 1383 * and mark a page dirty here because the caller is likely to mark 1384 * the page clean when we return, causing the page to possibly revert 1385 * to all-zero's later. 1386 */ 1387 } 1388 1389 static int 1390 swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, 1391 int *rbehind, int *rahead) 1392 { 1393 1394 VM_OBJECT_WLOCK(object); 1395 return (swap_pager_getpages_locked(object, ma, count, rbehind, rahead)); 1396 } 1397 1398 /* 1399 * swap_pager_getpages_async(): 1400 * 1401 * Right now this is emulation of asynchronous operation on top of 1402 * swap_pager_getpages(). 1403 */ 1404 static int 1405 swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, 1406 int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) 1407 { 1408 int r, error; 1409 1410 r = swap_pager_getpages(object, ma, count, rbehind, rahead); 1411 switch (r) { 1412 case VM_PAGER_OK: 1413 error = 0; 1414 break; 1415 case VM_PAGER_ERROR: 1416 error = EIO; 1417 break; 1418 case VM_PAGER_FAIL: 1419 error = EINVAL; 1420 break; 1421 default: 1422 panic("unhandled swap_pager_getpages() error %d", r); 1423 } 1424 (iodone)(arg, ma, count, error); 1425 1426 return (r); 1427 } 1428 1429 /* 1430 * swap_pager_putpages: 1431 * 1432 * Assign swap (if necessary) and initiate I/O on the specified pages. 1433 * 1434 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1435 * are automatically converted to SWAP objects. 1436 * 1437 * In a low memory situation we may block in VOP_STRATEGY(), but the new 1438 * vm_page reservation system coupled with properly written VFS devices 1439 * should ensure that no low-memory deadlock occurs. This is an area 1440 * which needs work. 1441 * 1442 * The parent has N vm_object_pip_add() references prior to 1443 * calling us and will remove references for rtvals[] that are 1444 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1445 * completion. 1446 * 1447 * The parent has soft-busy'd the pages it passes us and will unbusy 1448 * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. 1449 * We need to unbusy the rest on I/O completion. 1450 */ 1451 static void 1452 swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, 1453 int flags, int *rtvals) 1454 { 1455 struct buf *bp; 1456 daddr_t addr, blk, n_free, s_free; 1457 vm_page_t mreq; 1458 int i, j, n; 1459 bool async; 1460 1461 KASSERT(count == 0 || ma[0]->object == object, 1462 ("%s: object mismatch %p/%p", 1463 __func__, object, ma[0]->object)); 1464 1465 /* 1466 * Step 1 1467 * 1468 * Turn object into OBJT_SWAP. Force sync if not a pageout process. 1469 */ 1470 if (object->type != OBJT_SWAP) { 1471 addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); 1472 KASSERT(addr == SWAPBLK_NONE, 1473 ("unexpected object swap block")); 1474 } 1475 VM_OBJECT_WUNLOCK(object); 1476 async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; 1477 swp_pager_init_freerange(&s_free, &n_free); 1478 1479 /* 1480 * Step 2 1481 * 1482 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1483 * The page is left dirty until the pageout operation completes 1484 * successfully. 1485 */ 1486 for (i = 0; i < count; i += n) { 1487 /* Maximum I/O size is limited by maximum swap block size. */ 1488 n = min(count - i, nsw_cluster_max); 1489 1490 if (async) { 1491 mtx_lock(&swbuf_mtx); 1492 while (nsw_wcount_async == 0) 1493 msleep(&nsw_wcount_async, &swbuf_mtx, PVM, 1494 "swbufa", 0); 1495 nsw_wcount_async--; 1496 mtx_unlock(&swbuf_mtx); 1497 } 1498 1499 /* Get a block of swap of size up to size n. */ 1500 VM_OBJECT_WLOCK(object); 1501 blk = swp_pager_getswapspace(&n); 1502 if (blk == SWAPBLK_NONE) { 1503 VM_OBJECT_WUNLOCK(object); 1504 mtx_lock(&swbuf_mtx); 1505 if (++nsw_wcount_async == 1) 1506 wakeup(&nsw_wcount_async); 1507 mtx_unlock(&swbuf_mtx); 1508 for (j = 0; j < n; ++j) 1509 rtvals[i + j] = VM_PAGER_FAIL; 1510 continue; 1511 } 1512 for (j = 0; j < n; ++j) { 1513 mreq = ma[i + j]; 1514 vm_page_aflag_clear(mreq, PGA_SWAP_FREE); 1515 addr = swp_pager_meta_build(mreq->object, mreq->pindex, 1516 blk + j); 1517 if (addr != SWAPBLK_NONE) 1518 swp_pager_update_freerange(&s_free, &n_free, 1519 addr); 1520 MPASS(mreq->dirty == VM_PAGE_BITS_ALL); 1521 mreq->oflags |= VPO_SWAPINPROG; 1522 } 1523 VM_OBJECT_WUNLOCK(object); 1524 1525 bp = uma_zalloc(swwbuf_zone, M_WAITOK); 1526 if (async) 1527 bp->b_flags = B_ASYNC; 1528 bp->b_flags |= B_PAGING; 1529 bp->b_iocmd = BIO_WRITE; 1530 1531 bp->b_rcred = crhold(thread0.td_ucred); 1532 bp->b_wcred = crhold(thread0.td_ucred); 1533 bp->b_bcount = PAGE_SIZE * n; 1534 bp->b_bufsize = PAGE_SIZE * n; 1535 bp->b_blkno = blk; 1536 for (j = 0; j < n; j++) 1537 bp->b_pages[j] = ma[i + j]; 1538 bp->b_npages = n; 1539 1540 /* 1541 * Must set dirty range for NFS to work. 1542 */ 1543 bp->b_dirtyoff = 0; 1544 bp->b_dirtyend = bp->b_bcount; 1545 1546 VM_CNT_INC(v_swapout); 1547 VM_CNT_ADD(v_swappgsout, bp->b_npages); 1548 1549 /* 1550 * We unconditionally set rtvals[] to VM_PAGER_PEND so that we 1551 * can call the async completion routine at the end of a 1552 * synchronous I/O operation. Otherwise, our caller would 1553 * perform duplicate unbusy and wakeup operations on the page 1554 * and object, respectively. 1555 */ 1556 for (j = 0; j < n; j++) 1557 rtvals[i + j] = VM_PAGER_PEND; 1558 1559 /* 1560 * asynchronous 1561 * 1562 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1563 */ 1564 if (async) { 1565 bp->b_iodone = swp_pager_async_iodone; 1566 BUF_KERNPROC(bp); 1567 swp_pager_strategy(bp); 1568 continue; 1569 } 1570 1571 /* 1572 * synchronous 1573 * 1574 * NOTE: b_blkno is destroyed by the call to swapdev_strategy. 1575 */ 1576 bp->b_iodone = bdone; 1577 swp_pager_strategy(bp); 1578 1579 /* 1580 * Wait for the sync I/O to complete. 1581 */ 1582 bwait(bp, PVM, "swwrt"); 1583 1584 /* 1585 * Now that we are through with the bp, we can call the 1586 * normal async completion, which frees everything up. 1587 */ 1588 swp_pager_async_iodone(bp); 1589 } 1590 swp_pager_freeswapspace(s_free, n_free); 1591 VM_OBJECT_WLOCK(object); 1592 } 1593 1594 /* 1595 * swp_pager_async_iodone: 1596 * 1597 * Completion routine for asynchronous reads and writes from/to swap. 1598 * Also called manually by synchronous code to finish up a bp. 1599 * 1600 * This routine may not sleep. 1601 */ 1602 static void 1603 swp_pager_async_iodone(struct buf *bp) 1604 { 1605 int i; 1606 vm_object_t object = NULL; 1607 1608 /* 1609 * Report error - unless we ran out of memory, in which case 1610 * we've already logged it in swapgeom_strategy(). 1611 */ 1612 if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { 1613 printf( 1614 "swap_pager: I/O error - %s failed; blkno %ld," 1615 "size %ld, error %d\n", 1616 ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), 1617 (long)bp->b_blkno, 1618 (long)bp->b_bcount, 1619 bp->b_error 1620 ); 1621 } 1622 1623 /* 1624 * remove the mapping for kernel virtual 1625 */ 1626 if (buf_mapped(bp)) 1627 pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); 1628 else 1629 bp->b_data = bp->b_kvabase; 1630 1631 if (bp->b_npages) { 1632 object = bp->b_pages[0]->object; 1633 VM_OBJECT_WLOCK(object); 1634 } 1635 1636 /* 1637 * cleanup pages. If an error occurs writing to swap, we are in 1638 * very serious trouble. If it happens to be a disk error, though, 1639 * we may be able to recover by reassigning the swap later on. So 1640 * in this case we remove the m->swapblk assignment for the page 1641 * but do not free it in the rlist. The errornous block(s) are thus 1642 * never reallocated as swap. Redirty the page and continue. 1643 */ 1644 for (i = 0; i < bp->b_npages; ++i) { 1645 vm_page_t m = bp->b_pages[i]; 1646 1647 m->oflags &= ~VPO_SWAPINPROG; 1648 if (m->oflags & VPO_SWAPSLEEP) { 1649 m->oflags &= ~VPO_SWAPSLEEP; 1650 wakeup(&object->handle); 1651 } 1652 1653 /* We always have space after I/O, successful or not. */ 1654 vm_page_aflag_set(m, PGA_SWAP_SPACE); 1655 1656 if (bp->b_ioflags & BIO_ERROR) { 1657 /* 1658 * If an error occurs I'd love to throw the swapblk 1659 * away without freeing it back to swapspace, so it 1660 * can never be used again. But I can't from an 1661 * interrupt. 1662 */ 1663 if (bp->b_iocmd == BIO_READ) { 1664 /* 1665 * NOTE: for reads, m->dirty will probably 1666 * be overridden by the original caller of 1667 * getpages so don't play cute tricks here. 1668 */ 1669 vm_page_invalid(m); 1670 } else { 1671 /* 1672 * If a write error occurs, reactivate page 1673 * so it doesn't clog the inactive list, 1674 * then finish the I/O. 1675 */ 1676 MPASS(m->dirty == VM_PAGE_BITS_ALL); 1677 1678 /* PQ_UNSWAPPABLE? */ 1679 vm_page_activate(m); 1680 vm_page_sunbusy(m); 1681 } 1682 } else if (bp->b_iocmd == BIO_READ) { 1683 /* 1684 * NOTE: for reads, m->dirty will probably be 1685 * overridden by the original caller of getpages so 1686 * we cannot set them in order to free the underlying 1687 * swap in a low-swap situation. I don't think we'd 1688 * want to do that anyway, but it was an optimization 1689 * that existed in the old swapper for a time before 1690 * it got ripped out due to precisely this problem. 1691 */ 1692 KASSERT(!pmap_page_is_mapped(m), 1693 ("swp_pager_async_iodone: page %p is mapped", m)); 1694 KASSERT(m->dirty == 0, 1695 ("swp_pager_async_iodone: page %p is dirty", m)); 1696 1697 vm_page_valid(m); 1698 if (i < bp->b_pgbefore || 1699 i >= bp->b_npages - bp->b_pgafter) 1700 vm_page_readahead_finish(m); 1701 } else { 1702 /* 1703 * For write success, clear the dirty 1704 * status, then finish the I/O ( which decrements the 1705 * busy count and possibly wakes waiter's up ). 1706 * A page is only written to swap after a period of 1707 * inactivity. Therefore, we do not expect it to be 1708 * reused. 1709 */ 1710 KASSERT(!pmap_page_is_write_mapped(m), 1711 ("swp_pager_async_iodone: page %p is not write" 1712 " protected", m)); 1713 vm_page_undirty(m); 1714 vm_page_deactivate_noreuse(m); 1715 vm_page_sunbusy(m); 1716 } 1717 } 1718 1719 /* 1720 * adjust pip. NOTE: the original parent may still have its own 1721 * pip refs on the object. 1722 */ 1723 if (object != NULL) { 1724 vm_object_pip_wakeupn(object, bp->b_npages); 1725 VM_OBJECT_WUNLOCK(object); 1726 } 1727 1728 /* 1729 * swapdev_strategy() manually sets b_vp and b_bufobj before calling 1730 * bstrategy(). Set them back to NULL now we're done with it, or we'll 1731 * trigger a KASSERT in relpbuf(). 1732 */ 1733 if (bp->b_vp) { 1734 bp->b_vp = NULL; 1735 bp->b_bufobj = NULL; 1736 } 1737 /* 1738 * release the physical I/O buffer 1739 */ 1740 if (bp->b_flags & B_ASYNC) { 1741 mtx_lock(&swbuf_mtx); 1742 if (++nsw_wcount_async == 1) 1743 wakeup(&nsw_wcount_async); 1744 mtx_unlock(&swbuf_mtx); 1745 } 1746 uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); 1747 } 1748 1749 int 1750 swap_pager_nswapdev(void) 1751 { 1752 1753 return (nswapdev); 1754 } 1755 1756 static void 1757 swp_pager_force_dirty(vm_page_t m) 1758 { 1759 1760 vm_page_dirty(m); 1761 swap_pager_unswapped(m); 1762 vm_page_launder(m); 1763 } 1764 1765 /* 1766 * swap_pager_swapoff_object: 1767 * 1768 * Page in all of the pages that have been paged out for an object 1769 * to a swap device. 1770 */ 1771 static void 1772 swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) 1773 { 1774 struct swblk *sb; 1775 vm_page_t m; 1776 vm_pindex_t pi; 1777 daddr_t blk; 1778 int i, nv, rahead, rv; 1779 1780 KASSERT(object->type == OBJT_SWAP, 1781 ("%s: Object not swappable", __func__)); 1782 1783 for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 1784 &object->un_pager.swp.swp_blks, pi)) != NULL; ) { 1785 if ((object->flags & OBJ_DEAD) != 0) { 1786 /* 1787 * Make sure that pending writes finish before 1788 * returning. 1789 */ 1790 vm_object_pip_wait(object, "swpoff"); 1791 swp_pager_meta_free_all(object); 1792 break; 1793 } 1794 for (i = 0; i < SWAP_META_PAGES; i++) { 1795 /* 1796 * Count the number of contiguous valid blocks. 1797 */ 1798 for (nv = 0; nv < SWAP_META_PAGES - i; nv++) { 1799 blk = sb->d[i + nv]; 1800 if (!swp_pager_isondev(blk, sp) || 1801 blk == SWAPBLK_NONE) 1802 break; 1803 } 1804 if (nv == 0) 1805 continue; 1806 1807 /* 1808 * Look for a page corresponding to the first 1809 * valid block and ensure that any pending paging 1810 * operations on it are complete. If the page is valid, 1811 * mark it dirty and free the swap block. Try to batch 1812 * this operation since it may cause sp to be freed, 1813 * meaning that we must restart the scan. Avoid busying 1814 * valid pages since we may block forever on kernel 1815 * stack pages. 1816 */ 1817 m = vm_page_lookup(object, sb->p + i); 1818 if (m == NULL) { 1819 m = vm_page_alloc(object, sb->p + i, 1820 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); 1821 if (m == NULL) 1822 break; 1823 } else { 1824 if ((m->oflags & VPO_SWAPINPROG) != 0) { 1825 m->oflags |= VPO_SWAPSLEEP; 1826 VM_OBJECT_SLEEP(object, &object->handle, 1827 PSWP, "swpoff", 0); 1828 break; 1829 } 1830 if (vm_page_all_valid(m)) { 1831 do { 1832 swp_pager_force_dirty(m); 1833 } while (--nv > 0 && 1834 (m = vm_page_next(m)) != NULL && 1835 vm_page_all_valid(m) && 1836 (m->oflags & VPO_SWAPINPROG) == 0); 1837 break; 1838 } 1839 if (!vm_page_busy_acquire(m, VM_ALLOC_WAITFAIL)) 1840 break; 1841 } 1842 1843 vm_object_pip_add(object, 1); 1844 rahead = SWAP_META_PAGES; 1845 rv = swap_pager_getpages_locked(object, &m, 1, NULL, 1846 &rahead); 1847 if (rv != VM_PAGER_OK) 1848 panic("%s: read from swap failed: %d", 1849 __func__, rv); 1850 vm_object_pip_wakeupn(object, 1); 1851 VM_OBJECT_WLOCK(object); 1852 vm_page_xunbusy(m); 1853 1854 /* 1855 * The object lock was dropped so we must restart the 1856 * scan of this swap block. Pages paged in during this 1857 * iteration will be marked dirty in a future iteration. 1858 */ 1859 break; 1860 } 1861 if (i == SWAP_META_PAGES) 1862 pi = sb->p + SWAP_META_PAGES; 1863 } 1864 } 1865 1866 /* 1867 * swap_pager_swapoff: 1868 * 1869 * Page in all of the pages that have been paged out to the 1870 * given device. The corresponding blocks in the bitmap must be 1871 * marked as allocated and the device must be flagged SW_CLOSING. 1872 * There may be no processes swapped out to the device. 1873 * 1874 * This routine may block. 1875 */ 1876 static void 1877 swap_pager_swapoff(struct swdevt *sp) 1878 { 1879 vm_object_t object; 1880 int retries; 1881 1882 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 1883 1884 retries = 0; 1885 full_rescan: 1886 mtx_lock(&vm_object_list_mtx); 1887 TAILQ_FOREACH(object, &vm_object_list, object_list) { 1888 if (object->type != OBJT_SWAP) 1889 continue; 1890 mtx_unlock(&vm_object_list_mtx); 1891 /* Depends on type-stability. */ 1892 VM_OBJECT_WLOCK(object); 1893 1894 /* 1895 * Dead objects are eventually terminated on their own. 1896 */ 1897 if ((object->flags & OBJ_DEAD) != 0) 1898 goto next_obj; 1899 1900 /* 1901 * Sync with fences placed after pctrie 1902 * initialization. We must not access pctrie below 1903 * unless we checked that our object is swap and not 1904 * dead. 1905 */ 1906 atomic_thread_fence_acq(); 1907 if (object->type != OBJT_SWAP) 1908 goto next_obj; 1909 1910 swap_pager_swapoff_object(sp, object); 1911 next_obj: 1912 VM_OBJECT_WUNLOCK(object); 1913 mtx_lock(&vm_object_list_mtx); 1914 } 1915 mtx_unlock(&vm_object_list_mtx); 1916 1917 if (sp->sw_used) { 1918 /* 1919 * Objects may be locked or paging to the device being 1920 * removed, so we will miss their pages and need to 1921 * make another pass. We have marked this device as 1922 * SW_CLOSING, so the activity should finish soon. 1923 */ 1924 retries++; 1925 if (retries > 100) { 1926 panic("swapoff: failed to locate %d swap blocks", 1927 sp->sw_used); 1928 } 1929 pause("swpoff", hz / 20); 1930 goto full_rescan; 1931 } 1932 EVENTHANDLER_INVOKE(swapoff, sp); 1933 } 1934 1935 /************************************************************************ 1936 * SWAP META DATA * 1937 ************************************************************************ 1938 * 1939 * These routines manipulate the swap metadata stored in the 1940 * OBJT_SWAP object. 1941 * 1942 * Swap metadata is implemented with a global hash and not directly 1943 * linked into the object. Instead the object simply contains 1944 * appropriate tracking counters. 1945 */ 1946 1947 /* 1948 * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? 1949 */ 1950 static bool 1951 swp_pager_swblk_empty(struct swblk *sb, int start, int limit) 1952 { 1953 int i; 1954 1955 MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); 1956 for (i = start; i < limit; i++) { 1957 if (sb->d[i] != SWAPBLK_NONE) 1958 return (false); 1959 } 1960 return (true); 1961 } 1962 1963 /* 1964 * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free 1965 * 1966 * Nothing is done if the block is still in use. 1967 */ 1968 static void 1969 swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) 1970 { 1971 1972 if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { 1973 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 1974 uma_zfree(swblk_zone, sb); 1975 } 1976 } 1977 1978 /* 1979 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 1980 * 1981 * We first convert the object to a swap object if it is a default 1982 * object. 1983 * 1984 * The specified swapblk is added to the object's swap metadata. If 1985 * the swapblk is not valid, it is freed instead. Any previously 1986 * assigned swapblk is returned. 1987 */ 1988 static daddr_t 1989 swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) 1990 { 1991 static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; 1992 struct swblk *sb, *sb1; 1993 vm_pindex_t modpi, rdpi; 1994 daddr_t prev_swapblk; 1995 int error, i; 1996 1997 VM_OBJECT_ASSERT_WLOCKED(object); 1998 1999 /* 2000 * Convert default object to swap object if necessary 2001 */ 2002 if (object->type != OBJT_SWAP) { 2003 pctrie_init(&object->un_pager.swp.swp_blks); 2004 2005 /* 2006 * Ensure that swap_pager_swapoff()'s iteration over 2007 * object_list does not see a garbage pctrie. 2008 */ 2009 atomic_thread_fence_rel(); 2010 2011 object->type = OBJT_SWAP; 2012 object->un_pager.swp.writemappings = 0; 2013 KASSERT((object->flags & OBJ_ANON) != 0 || 2014 object->handle == NULL, 2015 ("default pager %p with handle %p", 2016 object, object->handle)); 2017 } 2018 2019 rdpi = rounddown(pindex, SWAP_META_PAGES); 2020 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); 2021 if (sb == NULL) { 2022 if (swapblk == SWAPBLK_NONE) 2023 return (SWAPBLK_NONE); 2024 for (;;) { 2025 sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == 2026 pageproc ? M_USE_RESERVE : 0)); 2027 if (sb != NULL) { 2028 sb->p = rdpi; 2029 for (i = 0; i < SWAP_META_PAGES; i++) 2030 sb->d[i] = SWAPBLK_NONE; 2031 if (atomic_cmpset_int(&swblk_zone_exhausted, 2032 1, 0)) 2033 printf("swblk zone ok\n"); 2034 break; 2035 } 2036 VM_OBJECT_WUNLOCK(object); 2037 if (uma_zone_exhausted(swblk_zone)) { 2038 if (atomic_cmpset_int(&swblk_zone_exhausted, 2039 0, 1)) 2040 printf("swap blk zone exhausted, " 2041 "increase kern.maxswzone\n"); 2042 vm_pageout_oom(VM_OOM_SWAPZ); 2043 pause("swzonxb", 10); 2044 } else 2045 uma_zwait(swblk_zone); 2046 VM_OBJECT_WLOCK(object); 2047 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2048 rdpi); 2049 if (sb != NULL) 2050 /* 2051 * Somebody swapped out a nearby page, 2052 * allocating swblk at the rdpi index, 2053 * while we dropped the object lock. 2054 */ 2055 goto allocated; 2056 } 2057 for (;;) { 2058 error = SWAP_PCTRIE_INSERT( 2059 &object->un_pager.swp.swp_blks, sb); 2060 if (error == 0) { 2061 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2062 1, 0)) 2063 printf("swpctrie zone ok\n"); 2064 break; 2065 } 2066 VM_OBJECT_WUNLOCK(object); 2067 if (uma_zone_exhausted(swpctrie_zone)) { 2068 if (atomic_cmpset_int(&swpctrie_zone_exhausted, 2069 0, 1)) 2070 printf("swap pctrie zone exhausted, " 2071 "increase kern.maxswzone\n"); 2072 vm_pageout_oom(VM_OOM_SWAPZ); 2073 pause("swzonxp", 10); 2074 } else 2075 uma_zwait(swpctrie_zone); 2076 VM_OBJECT_WLOCK(object); 2077 sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2078 rdpi); 2079 if (sb1 != NULL) { 2080 uma_zfree(swblk_zone, sb); 2081 sb = sb1; 2082 goto allocated; 2083 } 2084 } 2085 } 2086 allocated: 2087 MPASS(sb->p == rdpi); 2088 2089 modpi = pindex % SWAP_META_PAGES; 2090 /* Return prior contents of metadata. */ 2091 prev_swapblk = sb->d[modpi]; 2092 /* Enter block into metadata. */ 2093 sb->d[modpi] = swapblk; 2094 2095 /* 2096 * Free the swblk if we end up with the empty page run. 2097 */ 2098 if (swapblk == SWAPBLK_NONE) 2099 swp_pager_free_empty_swblk(object, sb); 2100 return (prev_swapblk); 2101 } 2102 2103 /* 2104 * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap 2105 * metadata, or transfer it into dstobject. 2106 * 2107 * This routine will free swap metadata structures as they are cleaned 2108 * out. 2109 */ 2110 static void 2111 swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, 2112 vm_pindex_t pindex, vm_pindex_t count) 2113 { 2114 struct swblk *sb; 2115 daddr_t n_free, s_free; 2116 vm_pindex_t offset, last; 2117 int i, limit, start; 2118 2119 VM_OBJECT_ASSERT_WLOCKED(srcobject); 2120 if (srcobject->type != OBJT_SWAP || count == 0) 2121 return; 2122 2123 swp_pager_init_freerange(&s_free, &n_free); 2124 offset = pindex; 2125 last = pindex + count; 2126 for (;;) { 2127 sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, 2128 rounddown(pindex, SWAP_META_PAGES)); 2129 if (sb == NULL || sb->p >= last) 2130 break; 2131 start = pindex > sb->p ? pindex - sb->p : 0; 2132 limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : 2133 SWAP_META_PAGES; 2134 for (i = start; i < limit; i++) { 2135 if (sb->d[i] == SWAPBLK_NONE) 2136 continue; 2137 if (dstobject == NULL || 2138 !swp_pager_xfer_source(srcobject, dstobject, 2139 sb->p + i - offset, sb->d[i])) { 2140 swp_pager_update_freerange(&s_free, &n_free, 2141 sb->d[i]); 2142 } 2143 sb->d[i] = SWAPBLK_NONE; 2144 } 2145 pindex = sb->p + SWAP_META_PAGES; 2146 if (swp_pager_swblk_empty(sb, 0, start) && 2147 swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { 2148 SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, 2149 sb->p); 2150 uma_zfree(swblk_zone, sb); 2151 } 2152 } 2153 swp_pager_freeswapspace(s_free, n_free); 2154 } 2155 2156 /* 2157 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2158 * 2159 * The requested range of blocks is freed, with any associated swap 2160 * returned to the swap bitmap. 2161 * 2162 * This routine will free swap metadata structures as they are cleaned 2163 * out. This routine does *NOT* operate on swap metadata associated 2164 * with resident pages. 2165 */ 2166 static void 2167 swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) 2168 { 2169 swp_pager_meta_transfer(object, NULL, pindex, count); 2170 } 2171 2172 /* 2173 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2174 * 2175 * This routine locates and destroys all swap metadata associated with 2176 * an object. 2177 */ 2178 static void 2179 swp_pager_meta_free_all(vm_object_t object) 2180 { 2181 struct swblk *sb; 2182 daddr_t n_free, s_free; 2183 vm_pindex_t pindex; 2184 int i; 2185 2186 VM_OBJECT_ASSERT_WLOCKED(object); 2187 if (object->type != OBJT_SWAP) 2188 return; 2189 2190 swp_pager_init_freerange(&s_free, &n_free); 2191 for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( 2192 &object->un_pager.swp.swp_blks, pindex)) != NULL;) { 2193 pindex = sb->p + SWAP_META_PAGES; 2194 for (i = 0; i < SWAP_META_PAGES; i++) { 2195 if (sb->d[i] == SWAPBLK_NONE) 2196 continue; 2197 swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); 2198 } 2199 SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); 2200 uma_zfree(swblk_zone, sb); 2201 } 2202 swp_pager_freeswapspace(s_free, n_free); 2203 } 2204 2205 /* 2206 * SWP_PAGER_METACTL() - misc control of swap meta data. 2207 * 2208 * This routine is capable of looking up, or removing swapblk 2209 * assignments in the swap meta data. It returns the swapblk being 2210 * looked-up, popped, or SWAPBLK_NONE if the block was invalid. 2211 * 2212 * When acting on a busy resident page and paging is in progress, we 2213 * have to wait until paging is complete but otherwise can act on the 2214 * busy page. 2215 */ 2216 static daddr_t 2217 swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) 2218 { 2219 struct swblk *sb; 2220 2221 VM_OBJECT_ASSERT_LOCKED(object); 2222 2223 /* 2224 * The meta data only exists if the object is OBJT_SWAP 2225 * and even then might not be allocated yet. 2226 */ 2227 KASSERT(object->type == OBJT_SWAP, 2228 ("Lookup object not swappable")); 2229 2230 sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, 2231 rounddown(pindex, SWAP_META_PAGES)); 2232 if (sb == NULL) 2233 return (SWAPBLK_NONE); 2234 return (sb->d[pindex % SWAP_META_PAGES]); 2235 } 2236 2237 /* 2238 * Returns the least page index which is greater than or equal to the 2239 * parameter pindex and for which there is a swap block allocated. 2240 * Returns object's size if the object's type is not swap or if there 2241 * are no allocated swap blocks for the object after the requested 2242 * pindex. 2243 */ 2244 vm_pindex_t 2245 swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) 2246 { 2247 struct swblk *sb; 2248 int i; 2249 2250 VM_OBJECT_ASSERT_LOCKED(object); 2251 if (object->type != OBJT_SWAP) 2252 return (object->size); 2253 2254 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2255 rounddown(pindex, SWAP_META_PAGES)); 2256 if (sb == NULL) 2257 return (object->size); 2258 if (sb->p < pindex) { 2259 for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { 2260 if (sb->d[i] != SWAPBLK_NONE) 2261 return (sb->p + i); 2262 } 2263 sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, 2264 roundup(pindex, SWAP_META_PAGES)); 2265 if (sb == NULL) 2266 return (object->size); 2267 } 2268 for (i = 0; i < SWAP_META_PAGES; i++) { 2269 if (sb->d[i] != SWAPBLK_NONE) 2270 return (sb->p + i); 2271 } 2272 2273 /* 2274 * We get here if a swblk is present in the trie but it 2275 * doesn't map any blocks. 2276 */ 2277 MPASS(0); 2278 return (object->size); 2279 } 2280 2281 /* 2282 * System call swapon(name) enables swapping on device name, 2283 * which must be in the swdevsw. Return EBUSY 2284 * if already swapping on this device. 2285 */ 2286 #ifndef _SYS_SYSPROTO_H_ 2287 struct swapon_args { 2288 char *name; 2289 }; 2290 #endif 2291 2292 /* 2293 * MPSAFE 2294 */ 2295 /* ARGSUSED */ 2296 int 2297 sys_swapon(struct thread *td, struct swapon_args *uap) 2298 { 2299 struct vattr attr; 2300 struct vnode *vp; 2301 struct nameidata nd; 2302 int error; 2303 2304 error = priv_check(td, PRIV_SWAPON); 2305 if (error) 2306 return (error); 2307 2308 sx_xlock(&swdev_syscall_lock); 2309 2310 /* 2311 * Swap metadata may not fit in the KVM if we have physical 2312 * memory of >1GB. 2313 */ 2314 if (swblk_zone == NULL) { 2315 error = ENOMEM; 2316 goto done; 2317 } 2318 2319 NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, 2320 uap->name, td); 2321 error = namei(&nd); 2322 if (error) 2323 goto done; 2324 2325 NDFREE(&nd, NDF_ONLY_PNBUF); 2326 vp = nd.ni_vp; 2327 2328 if (vn_isdisk(vp, &error)) { 2329 error = swapongeom(vp); 2330 } else if (vp->v_type == VREG && 2331 (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && 2332 (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { 2333 /* 2334 * Allow direct swapping to NFS regular files in the same 2335 * way that nfs_mountroot() sets up diskless swapping. 2336 */ 2337 error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); 2338 } 2339 2340 if (error) 2341 vrele(vp); 2342 done: 2343 sx_xunlock(&swdev_syscall_lock); 2344 return (error); 2345 } 2346 2347 /* 2348 * Check that the total amount of swap currently configured does not 2349 * exceed half the theoretical maximum. If it does, print a warning 2350 * message. 2351 */ 2352 static void 2353 swapon_check_swzone(void) 2354 { 2355 2356 /* recommend using no more than half that amount */ 2357 if (swap_total > swap_maxpages / 2) { 2358 printf("warning: total configured swap (%lu pages) " 2359 "exceeds maximum recommended amount (%lu pages).\n", 2360 swap_total, swap_maxpages / 2); 2361 printf("warning: increase kern.maxswzone " 2362 "or reduce amount of swap.\n"); 2363 } 2364 } 2365 2366 static void 2367 swaponsomething(struct vnode *vp, void *id, u_long nblks, 2368 sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) 2369 { 2370 struct swdevt *sp, *tsp; 2371 daddr_t dvbase; 2372 u_long mblocks; 2373 2374 /* 2375 * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. 2376 * First chop nblks off to page-align it, then convert. 2377 * 2378 * sw->sw_nblks is in page-sized chunks now too. 2379 */ 2380 nblks &= ~(ctodb(1) - 1); 2381 nblks = dbtoc(nblks); 2382 2383 /* 2384 * If we go beyond this, we get overflows in the radix 2385 * tree bitmap code. 2386 */ 2387 mblocks = 0x40000000 / BLIST_META_RADIX; 2388 if (nblks > mblocks) { 2389 printf( 2390 "WARNING: reducing swap size to maximum of %luMB per unit\n", 2391 mblocks / 1024 / 1024 * PAGE_SIZE); 2392 nblks = mblocks; 2393 } 2394 2395 sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); 2396 sp->sw_vp = vp; 2397 sp->sw_id = id; 2398 sp->sw_dev = dev; 2399 sp->sw_nblks = nblks; 2400 sp->sw_used = 0; 2401 sp->sw_strategy = strategy; 2402 sp->sw_close = close; 2403 sp->sw_flags = flags; 2404 2405 sp->sw_blist = blist_create(nblks, M_WAITOK); 2406 /* 2407 * Do not free the first blocks in order to avoid overwriting 2408 * any bsd label at the front of the partition 2409 */ 2410 blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), 2411 nblks - howmany(BBSIZE, PAGE_SIZE)); 2412 2413 dvbase = 0; 2414 mtx_lock(&sw_dev_mtx); 2415 TAILQ_FOREACH(tsp, &swtailq, sw_list) { 2416 if (tsp->sw_end >= dvbase) { 2417 /* 2418 * We put one uncovered page between the devices 2419 * in order to definitively prevent any cross-device 2420 * I/O requests 2421 */ 2422 dvbase = tsp->sw_end + 1; 2423 } 2424 } 2425 sp->sw_first = dvbase; 2426 sp->sw_end = dvbase + nblks; 2427 TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); 2428 nswapdev++; 2429 swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); 2430 swap_total += nblks; 2431 swapon_check_swzone(); 2432 swp_sizecheck(); 2433 mtx_unlock(&sw_dev_mtx); 2434 EVENTHANDLER_INVOKE(swapon, sp); 2435 } 2436 2437 /* 2438 * SYSCALL: swapoff(devname) 2439 * 2440 * Disable swapping on the given device. 2441 * 2442 * XXX: Badly designed system call: it should use a device index 2443 * rather than filename as specification. We keep sw_vp around 2444 * only to make this work. 2445 */ 2446 #ifndef _SYS_SYSPROTO_H_ 2447 struct swapoff_args { 2448 char *name; 2449 }; 2450 #endif 2451 2452 /* 2453 * MPSAFE 2454 */ 2455 /* ARGSUSED */ 2456 int 2457 sys_swapoff(struct thread *td, struct swapoff_args *uap) 2458 { 2459 struct vnode *vp; 2460 struct nameidata nd; 2461 struct swdevt *sp; 2462 int error; 2463 2464 error = priv_check(td, PRIV_SWAPOFF); 2465 if (error) 2466 return (error); 2467 2468 sx_xlock(&swdev_syscall_lock); 2469 2470 NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, 2471 td); 2472 error = namei(&nd); 2473 if (error) 2474 goto done; 2475 NDFREE(&nd, NDF_ONLY_PNBUF); 2476 vp = nd.ni_vp; 2477 2478 mtx_lock(&sw_dev_mtx); 2479 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2480 if (sp->sw_vp == vp) 2481 break; 2482 } 2483 mtx_unlock(&sw_dev_mtx); 2484 if (sp == NULL) { 2485 error = EINVAL; 2486 goto done; 2487 } 2488 error = swapoff_one(sp, td->td_ucred); 2489 done: 2490 sx_xunlock(&swdev_syscall_lock); 2491 return (error); 2492 } 2493 2494 static int 2495 swapoff_one(struct swdevt *sp, struct ucred *cred) 2496 { 2497 u_long nblks; 2498 #ifdef MAC 2499 int error; 2500 #endif 2501 2502 sx_assert(&swdev_syscall_lock, SA_XLOCKED); 2503 #ifdef MAC 2504 (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); 2505 error = mac_system_check_swapoff(cred, sp->sw_vp); 2506 (void) VOP_UNLOCK(sp->sw_vp); 2507 if (error != 0) 2508 return (error); 2509 #endif 2510 nblks = sp->sw_nblks; 2511 2512 /* 2513 * We can turn off this swap device safely only if the 2514 * available virtual memory in the system will fit the amount 2515 * of data we will have to page back in, plus an epsilon so 2516 * the system doesn't become critically low on swap space. 2517 */ 2518 if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) 2519 return (ENOMEM); 2520 2521 /* 2522 * Prevent further allocations on this device. 2523 */ 2524 mtx_lock(&sw_dev_mtx); 2525 sp->sw_flags |= SW_CLOSING; 2526 swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); 2527 swap_total -= nblks; 2528 mtx_unlock(&sw_dev_mtx); 2529 2530 /* 2531 * Page in the contents of the device and close it. 2532 */ 2533 swap_pager_swapoff(sp); 2534 2535 sp->sw_close(curthread, sp); 2536 mtx_lock(&sw_dev_mtx); 2537 sp->sw_id = NULL; 2538 TAILQ_REMOVE(&swtailq, sp, sw_list); 2539 nswapdev--; 2540 if (nswapdev == 0) { 2541 swap_pager_full = 2; 2542 swap_pager_almost_full = 1; 2543 } 2544 if (swdevhd == sp) 2545 swdevhd = NULL; 2546 mtx_unlock(&sw_dev_mtx); 2547 blist_destroy(sp->sw_blist); 2548 free(sp, M_VMPGDATA); 2549 return (0); 2550 } 2551 2552 void 2553 swapoff_all(void) 2554 { 2555 struct swdevt *sp, *spt; 2556 const char *devname; 2557 int error; 2558 2559 sx_xlock(&swdev_syscall_lock); 2560 2561 mtx_lock(&sw_dev_mtx); 2562 TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { 2563 mtx_unlock(&sw_dev_mtx); 2564 if (vn_isdisk(sp->sw_vp, NULL)) 2565 devname = devtoname(sp->sw_vp->v_rdev); 2566 else 2567 devname = "[file]"; 2568 error = swapoff_one(sp, thread0.td_ucred); 2569 if (error != 0) { 2570 printf("Cannot remove swap device %s (error=%d), " 2571 "skipping.\n", devname, error); 2572 } else if (bootverbose) { 2573 printf("Swap device %s removed.\n", devname); 2574 } 2575 mtx_lock(&sw_dev_mtx); 2576 } 2577 mtx_unlock(&sw_dev_mtx); 2578 2579 sx_xunlock(&swdev_syscall_lock); 2580 } 2581 2582 void 2583 swap_pager_status(int *total, int *used) 2584 { 2585 2586 *total = swap_total; 2587 *used = swap_total - swap_pager_avail - 2588 nswapdev * howmany(BBSIZE, PAGE_SIZE); 2589 } 2590 2591 int 2592 swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) 2593 { 2594 struct swdevt *sp; 2595 const char *tmp_devname; 2596 int error, n; 2597 2598 n = 0; 2599 error = ENOENT; 2600 mtx_lock(&sw_dev_mtx); 2601 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2602 if (n != name) { 2603 n++; 2604 continue; 2605 } 2606 xs->xsw_version = XSWDEV_VERSION; 2607 xs->xsw_dev = sp->sw_dev; 2608 xs->xsw_flags = sp->sw_flags; 2609 xs->xsw_nblks = sp->sw_nblks; 2610 xs->xsw_used = sp->sw_used; 2611 if (devname != NULL) { 2612 if (vn_isdisk(sp->sw_vp, NULL)) 2613 tmp_devname = devtoname(sp->sw_vp->v_rdev); 2614 else 2615 tmp_devname = "[file]"; 2616 strncpy(devname, tmp_devname, len); 2617 } 2618 error = 0; 2619 break; 2620 } 2621 mtx_unlock(&sw_dev_mtx); 2622 return (error); 2623 } 2624 2625 #if defined(COMPAT_FREEBSD11) 2626 #define XSWDEV_VERSION_11 1 2627 struct xswdev11 { 2628 u_int xsw_version; 2629 uint32_t xsw_dev; 2630 int xsw_flags; 2631 int xsw_nblks; 2632 int xsw_used; 2633 }; 2634 #endif 2635 2636 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2637 struct xswdev32 { 2638 u_int xsw_version; 2639 u_int xsw_dev1, xsw_dev2; 2640 int xsw_flags; 2641 int xsw_nblks; 2642 int xsw_used; 2643 }; 2644 #endif 2645 2646 static int 2647 sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) 2648 { 2649 struct xswdev xs; 2650 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2651 struct xswdev32 xs32; 2652 #endif 2653 #if defined(COMPAT_FREEBSD11) 2654 struct xswdev11 xs11; 2655 #endif 2656 int error; 2657 2658 if (arg2 != 1) /* name length */ 2659 return (EINVAL); 2660 error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); 2661 if (error != 0) 2662 return (error); 2663 #if defined(__amd64__) && defined(COMPAT_FREEBSD32) 2664 if (req->oldlen == sizeof(xs32)) { 2665 xs32.xsw_version = XSWDEV_VERSION; 2666 xs32.xsw_dev1 = xs.xsw_dev; 2667 xs32.xsw_dev2 = xs.xsw_dev >> 32; 2668 xs32.xsw_flags = xs.xsw_flags; 2669 xs32.xsw_nblks = xs.xsw_nblks; 2670 xs32.xsw_used = xs.xsw_used; 2671 error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); 2672 return (error); 2673 } 2674 #endif 2675 #if defined(COMPAT_FREEBSD11) 2676 if (req->oldlen == sizeof(xs11)) { 2677 xs11.xsw_version = XSWDEV_VERSION_11; 2678 xs11.xsw_dev = xs.xsw_dev; /* truncation */ 2679 xs11.xsw_flags = xs.xsw_flags; 2680 xs11.xsw_nblks = xs.xsw_nblks; 2681 xs11.xsw_used = xs.xsw_used; 2682 error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); 2683 return (error); 2684 } 2685 #endif 2686 error = SYSCTL_OUT(req, &xs, sizeof(xs)); 2687 return (error); 2688 } 2689 2690 SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, 2691 "Number of swap devices"); 2692 SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, 2693 sysctl_vm_swap_info, 2694 "Swap statistics by device"); 2695 2696 /* 2697 * Count the approximate swap usage in pages for a vmspace. The 2698 * shadowed or not yet copied on write swap blocks are not accounted. 2699 * The map must be locked. 2700 */ 2701 long 2702 vmspace_swap_count(struct vmspace *vmspace) 2703 { 2704 vm_map_t map; 2705 vm_map_entry_t cur; 2706 vm_object_t object; 2707 struct swblk *sb; 2708 vm_pindex_t e, pi; 2709 long count; 2710 int i; 2711 2712 map = &vmspace->vm_map; 2713 count = 0; 2714 2715 VM_MAP_ENTRY_FOREACH(cur, map) { 2716 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2717 continue; 2718 object = cur->object.vm_object; 2719 if (object == NULL || object->type != OBJT_SWAP) 2720 continue; 2721 VM_OBJECT_RLOCK(object); 2722 if (object->type != OBJT_SWAP) 2723 goto unlock; 2724 pi = OFF_TO_IDX(cur->offset); 2725 e = pi + OFF_TO_IDX(cur->end - cur->start); 2726 for (;; pi = sb->p + SWAP_META_PAGES) { 2727 sb = SWAP_PCTRIE_LOOKUP_GE( 2728 &object->un_pager.swp.swp_blks, pi); 2729 if (sb == NULL || sb->p >= e) 2730 break; 2731 for (i = 0; i < SWAP_META_PAGES; i++) { 2732 if (sb->p + i < e && 2733 sb->d[i] != SWAPBLK_NONE) 2734 count++; 2735 } 2736 } 2737 unlock: 2738 VM_OBJECT_RUNLOCK(object); 2739 } 2740 return (count); 2741 } 2742 2743 /* 2744 * GEOM backend 2745 * 2746 * Swapping onto disk devices. 2747 * 2748 */ 2749 2750 static g_orphan_t swapgeom_orphan; 2751 2752 static struct g_class g_swap_class = { 2753 .name = "SWAP", 2754 .version = G_VERSION, 2755 .orphan = swapgeom_orphan, 2756 }; 2757 2758 DECLARE_GEOM_CLASS(g_swap_class, g_class); 2759 2760 2761 static void 2762 swapgeom_close_ev(void *arg, int flags) 2763 { 2764 struct g_consumer *cp; 2765 2766 cp = arg; 2767 g_access(cp, -1, -1, 0); 2768 g_detach(cp); 2769 g_destroy_consumer(cp); 2770 } 2771 2772 /* 2773 * Add a reference to the g_consumer for an inflight transaction. 2774 */ 2775 static void 2776 swapgeom_acquire(struct g_consumer *cp) 2777 { 2778 2779 mtx_assert(&sw_dev_mtx, MA_OWNED); 2780 cp->index++; 2781 } 2782 2783 /* 2784 * Remove a reference from the g_consumer. Post a close event if all 2785 * references go away, since the function might be called from the 2786 * biodone context. 2787 */ 2788 static void 2789 swapgeom_release(struct g_consumer *cp, struct swdevt *sp) 2790 { 2791 2792 mtx_assert(&sw_dev_mtx, MA_OWNED); 2793 cp->index--; 2794 if (cp->index == 0) { 2795 if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) 2796 sp->sw_id = NULL; 2797 } 2798 } 2799 2800 static void 2801 swapgeom_done(struct bio *bp2) 2802 { 2803 struct swdevt *sp; 2804 struct buf *bp; 2805 struct g_consumer *cp; 2806 2807 bp = bp2->bio_caller2; 2808 cp = bp2->bio_from; 2809 bp->b_ioflags = bp2->bio_flags; 2810 if (bp2->bio_error) 2811 bp->b_ioflags |= BIO_ERROR; 2812 bp->b_resid = bp->b_bcount - bp2->bio_completed; 2813 bp->b_error = bp2->bio_error; 2814 bp->b_caller1 = NULL; 2815 bufdone(bp); 2816 sp = bp2->bio_caller1; 2817 mtx_lock(&sw_dev_mtx); 2818 swapgeom_release(cp, sp); 2819 mtx_unlock(&sw_dev_mtx); 2820 g_destroy_bio(bp2); 2821 } 2822 2823 static void 2824 swapgeom_strategy(struct buf *bp, struct swdevt *sp) 2825 { 2826 struct bio *bio; 2827 struct g_consumer *cp; 2828 2829 mtx_lock(&sw_dev_mtx); 2830 cp = sp->sw_id; 2831 if (cp == NULL) { 2832 mtx_unlock(&sw_dev_mtx); 2833 bp->b_error = ENXIO; 2834 bp->b_ioflags |= BIO_ERROR; 2835 bufdone(bp); 2836 return; 2837 } 2838 swapgeom_acquire(cp); 2839 mtx_unlock(&sw_dev_mtx); 2840 if (bp->b_iocmd == BIO_WRITE) 2841 bio = g_new_bio(); 2842 else 2843 bio = g_alloc_bio(); 2844 if (bio == NULL) { 2845 mtx_lock(&sw_dev_mtx); 2846 swapgeom_release(cp, sp); 2847 mtx_unlock(&sw_dev_mtx); 2848 bp->b_error = ENOMEM; 2849 bp->b_ioflags |= BIO_ERROR; 2850 printf("swap_pager: cannot allocate bio\n"); 2851 bufdone(bp); 2852 return; 2853 } 2854 2855 bp->b_caller1 = bio; 2856 bio->bio_caller1 = sp; 2857 bio->bio_caller2 = bp; 2858 bio->bio_cmd = bp->b_iocmd; 2859 bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; 2860 bio->bio_length = bp->b_bcount; 2861 bio->bio_done = swapgeom_done; 2862 if (!buf_mapped(bp)) { 2863 bio->bio_ma = bp->b_pages; 2864 bio->bio_data = unmapped_buf; 2865 bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 2866 bio->bio_ma_n = bp->b_npages; 2867 bio->bio_flags |= BIO_UNMAPPED; 2868 } else { 2869 bio->bio_data = bp->b_data; 2870 bio->bio_ma = NULL; 2871 } 2872 g_io_request(bio, cp); 2873 return; 2874 } 2875 2876 static void 2877 swapgeom_orphan(struct g_consumer *cp) 2878 { 2879 struct swdevt *sp; 2880 int destroy; 2881 2882 mtx_lock(&sw_dev_mtx); 2883 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2884 if (sp->sw_id == cp) { 2885 sp->sw_flags |= SW_CLOSING; 2886 break; 2887 } 2888 } 2889 /* 2890 * Drop reference we were created with. Do directly since we're in a 2891 * special context where we don't have to queue the call to 2892 * swapgeom_close_ev(). 2893 */ 2894 cp->index--; 2895 destroy = ((sp != NULL) && (cp->index == 0)); 2896 if (destroy) 2897 sp->sw_id = NULL; 2898 mtx_unlock(&sw_dev_mtx); 2899 if (destroy) 2900 swapgeom_close_ev(cp, 0); 2901 } 2902 2903 static void 2904 swapgeom_close(struct thread *td, struct swdevt *sw) 2905 { 2906 struct g_consumer *cp; 2907 2908 mtx_lock(&sw_dev_mtx); 2909 cp = sw->sw_id; 2910 sw->sw_id = NULL; 2911 mtx_unlock(&sw_dev_mtx); 2912 2913 /* 2914 * swapgeom_close() may be called from the biodone context, 2915 * where we cannot perform topology changes. Delegate the 2916 * work to the events thread. 2917 */ 2918 if (cp != NULL) 2919 g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); 2920 } 2921 2922 static int 2923 swapongeom_locked(struct cdev *dev, struct vnode *vp) 2924 { 2925 struct g_provider *pp; 2926 struct g_consumer *cp; 2927 static struct g_geom *gp; 2928 struct swdevt *sp; 2929 u_long nblks; 2930 int error; 2931 2932 pp = g_dev_getprovider(dev); 2933 if (pp == NULL) 2934 return (ENODEV); 2935 mtx_lock(&sw_dev_mtx); 2936 TAILQ_FOREACH(sp, &swtailq, sw_list) { 2937 cp = sp->sw_id; 2938 if (cp != NULL && cp->provider == pp) { 2939 mtx_unlock(&sw_dev_mtx); 2940 return (EBUSY); 2941 } 2942 } 2943 mtx_unlock(&sw_dev_mtx); 2944 if (gp == NULL) 2945 gp = g_new_geomf(&g_swap_class, "swap"); 2946 cp = g_new_consumer(gp); 2947 cp->index = 1; /* Number of active I/Os, plus one for being active. */ 2948 cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; 2949 g_attach(cp, pp); 2950 /* 2951 * XXX: Every time you think you can improve the margin for 2952 * footshooting, somebody depends on the ability to do so: 2953 * savecore(8) wants to write to our swapdev so we cannot 2954 * set an exclusive count :-( 2955 */ 2956 error = g_access(cp, 1, 1, 0); 2957 if (error != 0) { 2958 g_detach(cp); 2959 g_destroy_consumer(cp); 2960 return (error); 2961 } 2962 nblks = pp->mediasize / DEV_BSIZE; 2963 swaponsomething(vp, cp, nblks, swapgeom_strategy, 2964 swapgeom_close, dev2udev(dev), 2965 (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); 2966 return (0); 2967 } 2968 2969 static int 2970 swapongeom(struct vnode *vp) 2971 { 2972 int error; 2973 2974 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2975 if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { 2976 error = ENOENT; 2977 } else { 2978 g_topology_lock(); 2979 error = swapongeom_locked(vp->v_rdev, vp); 2980 g_topology_unlock(); 2981 } 2982 VOP_UNLOCK(vp); 2983 return (error); 2984 } 2985 2986 /* 2987 * VNODE backend 2988 * 2989 * This is used mainly for network filesystem (read: probably only tested 2990 * with NFS) swapfiles. 2991 * 2992 */ 2993 2994 static void 2995 swapdev_strategy(struct buf *bp, struct swdevt *sp) 2996 { 2997 struct vnode *vp2; 2998 2999 bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); 3000 3001 vp2 = sp->sw_id; 3002 vhold(vp2); 3003 if (bp->b_iocmd == BIO_WRITE) { 3004 if (bp->b_bufobj) 3005 bufobj_wdrop(bp->b_bufobj); 3006 bufobj_wref(&vp2->v_bufobj); 3007 } 3008 if (bp->b_bufobj != &vp2->v_bufobj) 3009 bp->b_bufobj = &vp2->v_bufobj; 3010 bp->b_vp = vp2; 3011 bp->b_iooffset = dbtob(bp->b_blkno); 3012 bstrategy(bp); 3013 return; 3014 } 3015 3016 static void 3017 swapdev_close(struct thread *td, struct swdevt *sp) 3018 { 3019 3020 VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); 3021 vrele(sp->sw_vp); 3022 } 3023 3024 3025 static int 3026 swaponvp(struct thread *td, struct vnode *vp, u_long nblks) 3027 { 3028 struct swdevt *sp; 3029 int error; 3030 3031 if (nblks == 0) 3032 return (ENXIO); 3033 mtx_lock(&sw_dev_mtx); 3034 TAILQ_FOREACH(sp, &swtailq, sw_list) { 3035 if (sp->sw_id == vp) { 3036 mtx_unlock(&sw_dev_mtx); 3037 return (EBUSY); 3038 } 3039 } 3040 mtx_unlock(&sw_dev_mtx); 3041 3042 (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 3043 #ifdef MAC 3044 error = mac_system_check_swapon(td->td_ucred, vp); 3045 if (error == 0) 3046 #endif 3047 error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); 3048 (void) VOP_UNLOCK(vp); 3049 if (error) 3050 return (error); 3051 3052 swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, 3053 NODEV, 0); 3054 return (0); 3055 } 3056 3057 static int 3058 sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) 3059 { 3060 int error, new, n; 3061 3062 new = nsw_wcount_async_max; 3063 error = sysctl_handle_int(oidp, &new, 0, req); 3064 if (error != 0 || req->newptr == NULL) 3065 return (error); 3066 3067 if (new > nswbuf / 2 || new < 1) 3068 return (EINVAL); 3069 3070 mtx_lock(&swbuf_mtx); 3071 while (nsw_wcount_async_max != new) { 3072 /* 3073 * Adjust difference. If the current async count is too low, 3074 * we will need to sqeeze our update slowly in. Sleep with a 3075 * higher priority than getpbuf() to finish faster. 3076 */ 3077 n = new - nsw_wcount_async_max; 3078 if (nsw_wcount_async + n >= 0) { 3079 nsw_wcount_async += n; 3080 nsw_wcount_async_max += n; 3081 wakeup(&nsw_wcount_async); 3082 } else { 3083 nsw_wcount_async_max -= nsw_wcount_async; 3084 nsw_wcount_async = 0; 3085 msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, 3086 "swpsysctl", 0); 3087 } 3088 } 3089 mtx_unlock(&swbuf_mtx); 3090 3091 return (0); 3092 } 3093 3094 static void 3095 swap_pager_update_writecount(vm_object_t object, vm_offset_t start, 3096 vm_offset_t end) 3097 { 3098 3099 VM_OBJECT_WLOCK(object); 3100 KASSERT((object->flags & OBJ_ANON) == 0, 3101 ("Splittable object with writecount")); 3102 object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; 3103 VM_OBJECT_WUNLOCK(object); 3104 } 3105 3106 static void 3107 swap_pager_release_writecount(vm_object_t object, vm_offset_t start, 3108 vm_offset_t end) 3109 { 3110 3111 VM_OBJECT_WLOCK(object); 3112 KASSERT((object->flags & OBJ_ANON) == 0, 3113 ("Splittable object with writecount")); 3114 object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; 3115 VM_OBJECT_WUNLOCK(object); 3116 } 3117